deep neural networks

Improving the quality of images in the field of computer vision has been raised as one of the keychallenges. In this paper, a new attention-based super-resolution adversarial generative neuralnetwork model is introduced, which is developed by combining real image resolution enhancementnetwork models and residual channel-based attention network to improve reconstruction of highresolutionimages and reduce complex distortions and noises. Image distortions include geometricdistortions, blurring, and loss of detail commonly seen in low-resolution or compressed images. Inthis model, the capabilities of the real-world image enhancement network to reduce noise andenhance clarity along with the ability of the residual channel-oriented attention network to preserveFine details are used. Experimental results on several examples of well-known data sets in this fieldhave shown that the proposed model improves the performance by about 5% on average in PIQEand . compared to existing methods such as attention enhancement network and real imageresolution enhancement network in well-known measures of blind image reconstruction. Theseimprovements have enabled the model to reconstruct images with high resolution and minimaldistortion and noise.

Deep neural networks (DNNs) have achieved great interest due to their success in various applications. However, the computation complexity and memory size are considered to be the main obstacles for implementing such models on embedded devices with limited memory and computational resources. Network compression techniques can overcome these challenges. Quantization and pruning methods are the most important compression techniques among them. One of the famous quantization methods in DNNs is the multi-level binary quantization, which not only exploits simple bit-wise logical operations, but also reduces the accuracy gap between binary neural networks and full precision DNNs. Since, multi-level binary can’t represent the zero value, this quantization does’nt take advantage of sparsity. On the other hand, it has been shown that DNNs are sparse, and by pruning the parameters of the DNNs, the amount of data storage in memory is reduced while computation speedup is also achieved.
In this paper, we propose a pruning and quantization-aware training method for multi-level ternary quantization that takes advantage of both multi-level quantization and data sparsity. In addition to increasing the accuracy of the network compared to the binary multi-level networks, it gives the network the ability to be sparse. To save memory size and computation complexity, we increase the sparsity in the quantized network by pruning until the accuracy loss is negligible. The results show that the potential speedup of computation for our model at the bit and word-level sparsity can be increased by 15x and 45x compared to the basic multi-level binary networks.

In recent decades, various techniques in the field of computer vision have been proposed to classify and identify scenes in different spaces on visible spectrum images. Recently, improvements in various areas of machine vision using complementary features in visible and infrared images have been reported by researchers. This approach can also work for the subject of scene recognition. In this paper, we first create a multispectral image database, including pairs of color-visible and infrared spectrum images. Then, the images are fused by wavelet transform and using a weighting method based on training the deep convolutional neural networks. This approach is also compared with several other fusion methods using quantitative evaluation criteria. Finally, images of different scenes are classified using architectures based on deeply trained convolutional neural networks. Transfer learning approach is also used to classify the scene with the least computational cost. Experimental results show that the proposed method has been effective in scene classification, which is done by fusion of four-channel images (RGB-IR) with a classification accuracy of 96.67%. This method also shows better quantitative fusion criteria than other multispectral image fusion approaches.